# Vision Zero Self Play Vlm Improvement

> Enable Vision-Language Models to improve without human annotations through competitive multi-agent self-play games (Who-Is-the-Spy format) combined with Iterative Self-Play Policy Optimization. Use when annotation costs limit multimodal dataset scaling or domain diversity.

- Skill: `adu2021/vision-zero-self-play-vlm-improvement` (Agent Skill)
- Install (CLI): `npx skillmds@latest add adu2021/vision-zero-self-play-vlm-improvement`
- Raw SKILL.md: https://api.skillmd.com/api/skills/adu2021/vision-zero-self-play-vlm-improvement/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: adu2021 (https://skillmd.com/u/adu2021)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/adu2021/vision-zero-self-play-vlm-improvement

---


# Vision-Zero: Label-Free Self-Play for VLM Self-Improvement

Vision-Zero addresses the bottleneck of expensive multimodal annotation by introducing a label-free framework where VLMs competitively play games (Who-Is-the-Spy) against each other, generating training signals without human labels while developing strategic reasoning and inference optimization.

## Core Architecture

- **Multi-agent game dynamics**: Asymmetric information game where spy identifies non-spy, non-spies identify spy
- **Self-play loop**: Agents alternate roles across episodes, learning both defensive and offensive visual reasoning
- **Iterative-SPO**: Policy optimization using game outcomes (win/loss) as sparse reward signals
- **Role-based training**: Asymmetric objectives shape distinct reasoning patterns for identification vs. deception

## Implementation Steps

Initialize self-play framework with VLM baseline:

```python
# Setup competitive game environment
from vision_zero import SelfPlayGame, IterativeSPO

# Create game with asymmetric roles
game = SelfPlayGame(
    model_name="Qwen2.5-VL-7B",
    num_agents=2,  # Spy vs. non-spies
    game_type="who_is_the_spy",
    max_rounds=5,
    image_domains=["synthetic", "chart", "real-world"]
)

# Initialize policy optimizer
optimizer = IterativeSPO(
    learning_rate=1e-5,
    policy_update_interval=100,  # update after every 100 games
    kl_coefficient=0.1,
    temperature=0.7
)
```

Execute self-play training with role switching:

```python
# Run self-play iterations
for iteration in range(num_iterations):
    # Episode with current policy
    trajectories = game.play_episodes(
        num_episodes=100,
        policy=current_policy,
        role_switch=True  # alternate spy/non-spy roles
    )

    # Compute rewards from game outcomes
    rewards = game.compute_rewards(trajectories)

    # Update policy using game signals
    current_policy = optimizer.update(trajectories, rewards)
```

## Practical Guidance

**When to use Vision-Zero:**
- Multimodal datasets with prohibitive annotation costs
- Domains where competitive reasoning improves visual understanding (e.g., forensic analysis, anomaly detection)
- VLM self-improvement without external human preference data
- Scenarios requiring diverse image understanding (synthetic, charts, photos)

**When NOT to use:**
- Tasks requiring supervised ground truth labels (e.g., medical imaging, legal documents)
- Single-player tasks lacking competitive framing
- Domain-specific reasoning without implicit competitive dynamics

**Hyperparameters:**
- **Game length (5 rounds)**: Increase to 7-10 for complex domains; decrease for rapid iteration
- **Temperature (0.7)**: Higher values (0.8-1.0) increase exploration; lower values focus on top-confident actions
- **KL coefficient (0.1)**: Increase to 0.2-0.3 to constrain policy drift from base model
- **Episode batch size (100)**: Scale proportionally to GPU memory; batch size affects gradient stability

## Cross-Domain Results

Vision-Zero demonstrates generalization across diverse visual domains:
- **Synthetic scenes (CLEVR)**: Precise geometric reasoning
- **Charts and diagrams**: Structured data extraction with comparative reasoning
- **Real-world images**: Open-ended visual understanding with uncertainty quantification

## References

Game-based multimodal learning extends prior work in reinforcement learning and curriculum design for language models.